Overview
DGX Spark vs Strix Halo comes down to this: both are small unified-memory boxes that hold models a 32 GB graphics card cannot, and the spec pages say the two have the same 256-bit LPDDR5x memory bus, so neither page gives a speed advantage you can quote. NVIDIA lists DGX Spark at 64 GB or 128 GB and 273 GB/s; AMD lists the Ryzen AI Max+ PRO 495 at up to 192 GB and no bandwidth. Tokens per second for either box is unknown on this page, because no vendor page read for it names a machine and a model with a speed. Sizes use the formulas on methodology.
DGX Spark vs Strix Halo: which is better for local LLMs?
Pick on memory size and software, not on a speed claim. Strix Halo is the codename AMD’s own article gives the Ryzen AI Max+395 (“codenamed Strix Halo”); AMD’s page for the Ryzen AI Max+ PRO 495 lists its former codename as Gorgon Halo, so by AMD’s own naming the 495 is not itself called Strix Halo; this page uses “Strix Halo” loosely for the Ryzen AI Max family. The spec rows side by side:
| Row | DGX Spark | Ryzen AI Max+ PRO 495 | RTX 5090 | RTX PRO 6000 Workstation |
|---|---|---|---|---|
| Memory | 64 GB LPDDR5X or 128 GB LPDDR5x, coherent unified | up to 192 GB, 256-bit LPDDR5x | 32 GB GDDR7 | 96 GB GDDR7 ECC |
| Interface | 256-bit | 256-bit | 512-bit | not listed |
| Bandwidth | 273 GB/s | not listed | not listed | 1792 GB/sec |
| Speed grade | not listed | LPDDR5x-8533 | not listed | not listed |
| Power | 240 W supply, GB10 TDP 140 W | default TDP 55 W, configurable 45-120 W | 575 W total graphics power | 600 W |
| Compute | up to 1 PFLOP FP4 | GPU: Radeon 8065S, 40 cores; NPU up to 55 TOPS | 3352 AI TOPS | 4000 AI TOPS, theoretical FP4 with sparsity |
Sources: NVIDIA DGX Spark, AMD Ryzen AI Max+ PRO 495, NVIDIA RTX 5090, NVIDIA RTX PRO 6000. The compute rows use different units and precisions (FP4 PFLOP, TOPS with sparsity), so they are not a like-for-like speed ranking. The AMD page lists the 495 as a PRO part (a business line) for laptops and desktops; the DGX Spark page lists the 64 GB option as available exclusively through participating OEM partners and marks it “Coming Soon”.
How much bandwidth does Strix Halo have?
AMD’s page lists no bandwidth, so this is a derived figure, not an AMD number. Bandwidth = bus width in bytes x transfer rate = (256 / 8) x 8533e6 = 32 x 8533e6 = 273.056 GB/s. That matches NVIDIA’s 273 GB/s for DGX Spark’s 256-bit LPDDR5x, which is what you would expect from the same bus width and a similar LPDDR5x speed grade. NVIDIA’s page does not list DGX Spark’s transfer rate, so the match is arithmetic, not a vendor statement that the memory is the same. It also assumes the part runs at its maximum listed speed, and the actual speed of a given machine is unknown.
DGX Spark vs RTX 5090 and RTX PRO 6000: when is the GPU better?
A discrete GPU has far more memory bandwidth per byte of capacity but far less capacity. The RTX PRO 6000 Workstation lists 1792 GB/sec, which is 1792 / 273 = 6.56 times DGX Spark’s 273 GB/s. NVIDIA’s RTX 5090 page lists a 512-bit bus but no bandwidth, so that ratio is unknown for the 5090. Because generating text reads the weights from memory for every token, bandwidth is the main thing that separates these boxes once a model fits, but this page does not turn that into a tokens-per-second number, because none is published here. The trade is capacity against bandwidth: 32 GB on a 5090 (rtx-5090-llm), 96 GB on the PRO 6000 (rtx-5090-vs-rtx-pro-6000), 64 or 128 GB on DGX Spark. Prices are not quoted; check current retailer price.
What does a 70B model need, and which box holds it?
Use the site formula: weights_GB = params x bits / 8 / 1e9, plus KV cache, plus a labelled 1 to 2 GB overhead assumption (vram-for-70b-model). The llama.cpp quantize README lists a 70B model at 43.1 GB for Q4_K_M. KV cache is 2 x layers x kv_heads x head_dim x context x bytes per value. The Qwen2.5-72B config.json gives 80 layers, 8 KV heads, hidden size 8192 over 64 heads (head dimension 128) and bfloat16 values, so at 4,096 tokens KV = 2 x 80 x 8 x 128 x 4096 x 2 = 1.3422 GB.
Total = 43.1 + 1.3422 + (1 to 2) = 45.44 to 46.44 GB (README file size for the weights, Qwen config for the KV term; an approximation across two model families).
| Memory | Holds 45.44 to 46.44 GB at 4,096 context? |
|---|---|
| RTX 5090, 32 GB | No: the weights alone (43.1 GB) exceed 32 GB |
| DGX Spark 64 GB | Yes on capacity, with about 17.6 to 18.6 GB left (64 - 46.44 to 64 - 45.44) |
| DGX Spark 128 GB | Yes, about 81.6 to 82.6 GB left |
| RTX PRO 6000, 96 GB | Yes, about 49.6 to 50.6 GB left |
| Ryzen AI Max+ 395, 128 GB | Yes on capacity; AMD says up to 112 GB is allocatable to the GPU |
The leftover figures ignore the operating system, which also uses unified memory on the two boxes. On NVIDIA’s page, the 64 GB Spark is rated for models “up to 100 billion parameters” and the 128 GB for “up to 200 billion parameters”; those are NVIDIA’s claims, not measurements here. See ram-vs-vram for why shared memory behaves differently from VRAM.
What does the 64 GB DGX Spark option change?
It halves the capacity at the same listed 273 GB/s bus, and NVIDIA’s page rates a single 64 GB unit for up to 100B-parameter models against 200B for the 128 GB unit. A 70B at Q4_K_M fits in 64 GB on the math above, but with a 17.6 to 18.6 GB margin, so long contexts and a desktop on the same memory leave little room. Whether a 64 GB unit costs less or ships in your region is unknown here; the page says only that it comes through participating OEM partners.
Can the Ryzen AI Max+ PRO 495 use 192 GB for an LLM?
The AMD page lists 192 GB as the maximum memory, not how much a GPU can address. The only allocation figure AMD gives in the sources read is for the older Ryzen AI Max+395: 128 GB of unified memory with “up to 112 GB allocatable by the GPU”. For the 495 at 192 GB, the GPU-allocatable amount is unknown. Whether any laptop or mini PC ships with 192 GB is also unknown from these pages. AMD’s 395 article names a 70B model run (DeepSeek-R1-Distill-Llama-70B) on a 128 GB machine, but gives only relative speedups (for example up to 2.6x faster token generation in a concurrent test against a MacBook Pro) and no absolute tokens-per-second figure, so there is nothing to quote. See amd-gpu-for-local-llm for AMD discrete cards.
Is a unified-memory box better than a GPU for big models?
For fitting a model, yes: 64 to 192 GB of listed memory against 32 GB. For speed on a model that fits both, the discrete GPU has the higher listed bandwidth where a bandwidth is published (1792 vs 273 GB/s), and the unified boxes win only where the GPU would have to spill to system RAM. Mixture-of-experts models change this because only some experts are read per token; that case is on moe-cpu-offload, and big-model sizes are on deepseek-vram-requirements. A Mac is the other unified-memory route: which Mac for a local LLM.
How does DGX Station compare with DGX Spark?
NVIDIA’s DGX Station page lists the GB300 Grace Blackwell Ultra Desktop Superchip, 252 GB HBM3e at 7.1 TB/s, 748 GB of coherent memory, 396 GB/s for the 496 GB LPDDR5X, and up to 20 petaFLOPS. It is a different class from Spark’s 273 GB/s and up to 1 PFLOP FP4. When read, the page said to sign up to be notified when it becomes available, so treat it as not yet purchasable. Price is unknown.
Questions
DGX Spark vs Strix Halo: which is better for local LLMs?
Neither is shown to be faster on the vendor pages read, because no tokens-per-second figure with a named machine is published for either. DGX Spark lists 64 or 128 GB at 273 GB/s; the Ryzen AI Max+ PRO 495 lists up to 192 GB on a 256-bit LPDDR5x-8533 bus, which derives to 273.056 GB/s. Choose on capacity, software and price (check current retailer price).
DGX Spark vs RTX 5090 for local LLMs?
The RTX 5090 has 32 GB GDDR7 and cannot hold a 70B at Q4_K_M, whose 43.1 GB weights alone exceed it. DGX Spark holds it with 64 or 128 GB of unified memory. The 5090’s bandwidth is not listed on NVIDIA’s page, so the speed comparison is unknown.
Is a unified-memory box better than a GPU for big models?
It is better at fitting them and not shown to be better at running them fast. The RTX PRO 6000 lists 1792 GB/sec, 6.56 times Spark’s 273 GB/s, while Spark lists more memory per box than a 5090. Use the fit table above for capacity.
What does the 64 GB DGX Spark option change?
Capacity halves from 128 GB to 64 GB at the same listed 273 GB/s, and NVIDIA’s rated model size drops from 200B to 100B parameters. A 70B Q4_K_M plus cache is 45.44 to 46.44 GB and still fits, with little margin.
Can the Ryzen AI Max+ PRO 495 use 192 GB for an LLM?
192 GB is the listed maximum memory, but the GPU-allocatable share for the 495 is unknown. For the Ryzen AI Max+395 AMD states up to 112 GB of 128 GB is allocatable to the GPU.
Why does Ollama feel slow on DGX Spark?
Unknown. No primary source read for this page measures Ollama on DGX Spark, so any cause would be a guess. The measurable part is the 273 GB/s memory bus, which is lower than the RTX PRO 6000’s listed 1792 GB/sec. For general runtime behavior see llama-cpp-vs-ollama-vs-vllm and, if a GPU is not used at all, ollama-not-using-gpu.
Sources read on 2026-10-13
- NVIDIA DGX Spark: 64 GB / 128 GB LPDDR5x, 256-bit, 273 GB/s, GB10, 1 PFLOP FP4, 240 W supply, 140 W GB10 TDP, 64 GB via OEM partners, model-size claims.
- NVIDIA DGX Station: GB300, 252 GB HBM3e at 7.1 TB/s, 748 GB coherent, 396 GB/s CPU memory, 20 petaFLOPS, not yet available.
- AMD Ryzen AI Max+ PRO 495: 256-bit LPDDR5x, 192 GB max, LPDDR5x-8533, Radeon 8065S with 40 cores, NPU up to 55 TOPS, TDP 55 W (45-120 W configurable). No bandwidth listed.
- AMD Ryzen AI Max+ 395 technical article: “codenamed Strix Halo”, 128 GB with up to 112 GB allocatable by the GPU, 70B model run named.
- NVIDIA RTX 5090: 32 GB GDDR7, 512-bit, 3352 AI TOPS, 575 W; no bandwidth listed.
- NVIDIA RTX PRO 6000: 96 GB GDDR7 ECC, 1792 GB/sec, 600 W, 4000 AI TOPS (FP4 with sparsity); Workstation Edition only.
- llama.cpp quantize README: 70B Q4_K_M at 43.1 GB.
- Qwen2.5-72B config.json: 80 layers, 8 KV heads, hidden size 8192, 64 heads, bfloat16.
Related
- rtx-5090-llm and rtx-5090-vs-rtx-pro-6000 cover the GPU side.
- amd-gpu-for-local-llm covers AMD discrete cards.
- ram-vs-vram, vram-for-70b-model and deepseek-vram-requirements cover memory and big-model sizes.
- moe-cpu-offload, llama-cpp-vs-ollama-vs-vllm and ollama-not-using-gpu cover runtimes.
- methodology is the citation rule.
- Mac route: which Mac for a local LLM.